Controlled Diversity: Length-optimized Natural Language Generation

  • LLMs are not generally able to adjust the length of their outputs based on strict length requirements, a capability that would improve their usefulness in applications that require adherence to diverse user and system requirements. We present an approach to train LLMs to acquire this capability by augmenting existing data and applying existing fine-tuning techniques, which we compare based on theLLMs are not generally able to adjust the length of their outputs based on strict length requirements, a capability that would improve their usefulness in applications that require adherence to diverse user and system requirements. We present an approach to train LLMs to acquire this capability by augmenting existing data and applying existing fine-tuning techniques, which we compare based on the trained models’ adherence to the length requirement and overall response quality relative to the baseline model. Our results demonstrate that these techniques can be successfully applied to train LLMs to adhere to length requirements, with the trained models generating texts which better align to the length requirements. Our results indicate that our method may change the response quality when using training data that was not generated by the baseline model. This allows simultaneous alignment to another training objective in certain scenarios, but is undesirable otherwise. Training on a dataset containing the model’s own responses eliminates this issue.show moreshow less

Download full text files

Export metadata

Additional Services

Share in Twitter Search Google Scholar Statistics
Metadaten
Author:Diana Marie Schenke, Timo BaumannORCiDGND
URN:urn:nbn:de:bvb:898-opus4-80534
DOI:https://doi.org/10.48550/arXiv.2502.19347
ArXiv Id:http://arxiv.org/abs/arXiv:2502.19347
Parent Title (English):Proceedings of the ISCA/ITG Workshop on Diversity in Large Speech and Language Models, February 20, 2025 in Berlin, Germany
Editor:Sebastian Möller, Pia Knoeferle, Britta Schulte, Nils Feldhus
Document Type:conference proceeding (article)
Language:English
Year of first Publication:2025
Publishing Institution:Ostbayerische Technische Hochschule Regensburg
Release Date:2025/04/28
Pagenumber:8
Institutes:Fakultät Informatik und Mathematik
Begutachtungsstatus:peer-reviewed
research focus:Digitale Transformation
Licence (German):Creative Commons - CC BY - Namensnennung 4.0 International
Einverstanden ✔
Diese Webseite verwendet technisch erforderliche Session-Cookies. Durch die weitere Nutzung der Webseite stimmen Sie diesem zu. Unsere Datenschutzerklärung finden Sie hier.